How Dogs Learn Words
By neuroimaging dogs in an fMRI scanner, we see the canine origins of learning.
Posted Oct 31, 2018
What makes a dog able to learn? New research suggests that it may be the gift of confusion.
Let me unpack that.
My dog, Samson, is about 1.5 years old. Like most dogs, he knows that words mean things. He knows that there are words like ball and walk and yes, and he knows what they mean. Of course, this is nothing compared to Chaser, a famous border collie, who knew more than a thousand words and could retrieve objects based on their label.
But what Samson and Chaser have also learned is something far more important than the labels of any individual objects: They've learned that a sound can refer to things (like balls), verbs (like walking), and to affirmations (yes, keep doing that). And that means that new sounds can refer to new things.
The border collie Rico did something even more amazing to demonstrate this. He could recognize that a new word, one he hadn't heard before, went with a new object. When showing Rico two objects — one novel and one for which he already knew the name — he was more likely to select the novel object if he heard a novel name. Toddlers also do this; developmental psychologists call it the principle of mutual exclusivity.
Recently, dogs have been trained to lie still during fMRI scans. This has allowed researchers to figure out what parts of the dog brain are associated with reward, face detection, and smells.
In a recent study by Prichard et al. (2018), they had owners train their dogs to discriminate between two objects, like a monkey and a duck. Once the dogs could reliably discriminate between these objects, they had the dogs enter into a dog-friendly fMRI scanner, and they watched each dog's brain as it responded to the familiar words for the two objects, as well as two pseudo-words the dogs had never heard before.
The results showed that the dogs responded differently to known words than they did to pseudo-words. In particular, the pseudo-words showed substantially more activation than the known words, with most of this activation taking place in the auditory cortex and nearby regions of the parietotemporal area.
This is different from humans observed in a similar experiment. Humans show more activation for known words, with most of this activation in regions that are collectively referred to as the general semantic network (middle temporal gyrus, posterior inferior parietal lobe, fusiform and parahippocampal gyri, dorsomedial prefrontal cortex, and posterior cingulate gyrus).
One might be tempted to say that since dogs don't show a pattern of activation similar to the human pattern of the general semantic network, they don't have semantics. But that would be like saying my computer doesn't have semantics, because it also doesn't show a similar pattern of activation. There's no point in imaging my computer, because it doesn't have a brain. Yet it does have semantics, because I've programmed it to have them. Semantics doesn't require a specific brain area.
The authors knowingly steer clear of this reverse inference problem. (For more on reverse inverse problems in imaging, see Poldrack, 2011.) Instead, they suggest that the parietotemporal region seems associated with novel word detection in dogs. This is a tricky inference, but is true enough. However, it's important to remember that repeated exposure to a stimulus leads to what is called repetition suppression. Areas that are initially active become less active with repeated exposure.
Some researchers suggest this has to do with a phenomenon known as sharpening (Grill-spector et al, 2006). Sharpening is the brain getting better at differentiating what matters from what doesn't matter. When you first hear a weird signal, you might look around, paying attention to a lot of things as you try to figure out what it is. But later on, once you realize it's your new phone ringtone, your brain shows less activation the next time it hears the (now not-so-weird) signal.
The more familiar the signal, the less brain activation. This pattern of general to specific activation is similar to what deep neural networks do during training: They often start off by overgeneralizing and then sharpening.
This overgeneralization is a form of confusion. Indeed, it might be the feeling of confusion. There is more activation everywhere, as the brain (or neural network) tries to sort out what matters from what doesn't matter. So it may be that the activation in the parietotemporal lobe is really signaling the dog's confusion.
This is beautiful, because it suggests that the dogs know that there is something to know, they just don't know what it is. They get confused — and that makes it possible for them to learn.
This line of reasoning would also explain why people don't show a similar pattern. If you show up for an experiment, and someone speaks perfect English to you except in this one situation where they say a nonsense word, chances are you're going to believe that word is nonsense, which in fact it is. But dogs probably think we're speaking nonsense most of the time, except when we're making eye contact and saying the same word repeatedly, which is what the owners did in the experiment.
Consider that idea in reverse. If a person thinks they know everything, chances are, they are not going to learn very much. They will rarely be confused, because they have a neat set of packages that explain everything (even the stuff they've never heard of). If they know everything, then they already know why we die, why Jamal Khashoggi is important, and that algorithms can have mental health problems. And if they know everything but don't know that stuff, then they know that stuff isn't important.
On the other hand, if they recognize that there are things they don't know, then they have the capacity for confusion. And confusion is what makes the brain figure things out.
In the land of dogs, the ability of a dog to be confused is what makes dogs able to learn.
Facebook Image: VK Studio/Shutterstock
Prichard, A., Cook, P. F., Spivak, M., Chhibber, R., & Berns, G. (2018). Awake fMRI Reveals Brain Regions for Novel Word Detection in Dogs. bioRxiv, 178186.
Poldrack, R. A. (2011). Inferring mental states from neuroimaging data: from reverse inference to large-scale decoding. Neuron, 72(5), 692-697.
Grill-Spector, K., Henson, R., & Martin, A. (2006). Repetition and the brain: neural models of stimulus-specific effects. Trends in cognitive sciences, 10(1), 14-23.